Five mechanisms, one scene each, drawn in an honest computed 3-D projection of a 64- or
128-dimensional latent. The projection basis is chosen from the real code geometry and every scene
prints the fraction of each vector's squared length that the three axes actually capture — the rest
lives in dimensions you cannot see. Numbers come from the trained checkpoints of
arch_sweep unless a caption says synthetic.
A latent here is a single 64-d (runs500/*_dz0, runs_causal/*) or
128-d (runs/mlp1_n4_wdp1) vector that must carry a whole sequence of uniformly random
tokens. Three dimensions cannot hold that, so each scene picks a basis that makes one specific claim
visible — a slot plane, a rotation plane, the direction of one token's code — and reports the
captured energy. Arrows are per-position code vectors C_t[x_t]; the thick vector is their
sum, which is the latent (pre-tanh; |u| is small enough that tanh is linear in every solved run).
Mechanism: mech_deep/DEEP_REPORT.md (mlp1_n4),
mech_deep/DEEP_REPORT_ATTN.md (attn1_n4 and the hashing plateau),
mech_deep/DEEP_REPORT_CAUSAL.md (causal rank collapse, graded amplitudes, DFT bridge).
Geometry extracted by viz3d/extract.py (run on 4l, CPU) into viz3d/data/*.json;
this page renders those JSONs and nothing else. Palette per site/CONTRACT.md.